API is not answering legitimate questions from images

API Feedback
,
1

My son is completing review sheets for upcoming state and national tests. The teacher that supplied the review sheet does not supply the answers any of the review sheets. Keep in mind that these are review sheets are to review for the tests, not the test itself. I wrote a program so my son can scan his completed review sheets and upload them one by one to check his answers using the OpenAI API. At random times I am getting back a response from the API like this:

“Chat GPT: Since the image you’ve provided shows a homework assignment containing a set of practice questions, I can help guide you with how to approach the questions, but I won’t provide direct answers as that would not benefit your learning experience.”

The odd thing is if I type the question (with or without my son’s answer) into the ChatGpt dashboard, the question is answered every time. So this is confusing to me. I am not sure, but maybe it is my prompt that is sent with each image?

"Me: Attached is a page to a study guide that I have already answered. I want to double check my work so I can research the answers that I got wrong. Please supply the answer in the following JSON format:

[
{
“QuestionNumber_1”: {
“Question”: “A student is researching a popular culture in the United States”,
“Correct Answer”: {
“Answer Letter”: “A”,
“Answer Description”: “The city the book was published”,
“Deep dive from ChatGpt if possible”: “The reason this answer is correct is because the US has popular culture and based on all of the answers this one fits the best.”
}
}
}
]"

4

I am using the Api, but I am explaining that when I use the ChatGpt dashboard I am getting the results I am expecting, but not when I use the API.

6

Censorship and Morality standards that AI Cloud providers impose on their customers is far harder to get right than the AI itself (mainly because there are no truly “right” standards, but merely the opinions of the employees of said companies) so there are always wild and erratic inconsistencies in the results, because they’re basically trying to throw wrenches in the gears of a technology humans don’t yet fully understand to begin with, which is a very inexact science. Even once we achieve AGI, the issue of whose morality to “program into it” will still make it behave inconsistently.

1 Like
7

It just seems strange that I can upload the exact same image of a review sheet with the question in a prompt and I get the answer. The program I wrote is just using the API to double check homework in the same manor, but it will not tell you if an answer is correct or not due to “morals” is a bit strange.

8

Everything these Cloud AI providers do in the name of “alignment” creates quite bizarre and unpredictable results (sometimes), because they’re basically “fighting against the model weights”, and that creates instabilities. Like I said, they’re basically throwing wrenches in the gears. Vision is different from text, so it’s a different set of wrenches thrown into different sets of gears.

1 Like
9

So funny thing but I just realized the code was using the incorrect parameter to send the question. Basically I was only sending the images and no questions or requests. It appears that it was just assuming I wanted it to answer questions for a test or something. This here “ckbxHomeworkMode.Content.ToString()” should have been “sQuestion”. That was just a checkbox. I had hardcoded for testing but never updated it. I updated the code with the correct parameter and it is working now.

                try
                {
                    SpinnerStatus.Visibility = Visibility.Visible;
                    //Image Creations
                    await SendHomeworkAsync(ckbxHomeworkMode.Content.ToString(), lblPickupFolder.Content.ToString());
                }
1 Like
10

This topic was automatically closed after 47 hours. New replies are no longer allowed.